Claude Fable 5.1 behaves differently without a code change. Here is what I hand it now

Disclosure: this piece was drafted with Claude Fable 5.1 on its launch day, inside the pipeline that writes every SharpHaw post. I read it, corrected it and edited it before it went out.

Anthropic shipped Claude Fable 5.1 on 1 September. The benchmarks are in every feed already. What is more useful to a developer who also runs a business is the page under the benchmarks: the release notes list what got better, and then, further down, what now behaves differently without you changing a line. I read both lists and sorted my own work into what I hand the model and what I keep. Here is the list, with the reasons taken from the notes rather than from vibes.

What changed without a code change

These are documented behaviour differences from Fable 5, not my guesses:

• Parallel tool calling is more variable. In coding and computer-use loops where the next reads are only implied, it may issue one tool call per turn. Same answer quality, more round trips.
• Fewer progress updates during long tool runs, especially at higher effort. Your UI can go quiet for minutes.
• At low effort it answers from memory more often and calls search and retrieval tools less.
• Denser prose in places, and less formatting in chat.
• When summarising documents it is more likely to reproduce passages without marking them as quotations.
• It is more likely to rewrite a whole file for a small change. Two things are breaking if you already call Fable 5. Forced tool use is gone: tool_choice set to any or a named tool returns a 400. And thinking blocks are now bound to the conversation that produced them. For accounts created on or after 31 August 2026, replaying a thinking block after you edited anything before it (system prompt, tool list, an earlier message) returns a 400 unless you opt into dropping the block. Future models are expected to enforce this for everyone.

Pricing is unchanged on input and output. Cache reads dropped to a quarter of the old price, which Anthropic puts at roughly 25% cheaper for typical work and up to 45% for agentic loops that keep re-reading the same context.

The hand list

I sort jobs by one question: what does a wrong answer cost, and who catches it first. Everything below ends in a draft, a read, or a tool only I use.

The years-old bug. Anthropic's launch post has Millennium saying it traced a one-in-a-million crash to a bug in a vendor library after their engineers had chased it for years. The notes say the model now fixes root causes rather than symptoms. The check is the same as any other patch: a failing test first, then the fix, then a human reads the diff.

The long refactor or migration. Multi-file changes across sessions that run for hours are the headline improvement. Let it run unattended, but give it a repository you own and a branch you can throw away. The MongoDB quote in the launch post is the honest picture: the engineer woke up to the next phase finished with a visual walkthrough, and then reviewed it.

The sourced research memo. Competitor pricing, a regulation, a library's actual behaviour. Run it at high effort with search on and ask for the source next to every claim. Then open three at random. If one does not say what the memo says, the memo goes back.

The document read. Contracts, terms of service, the fifty-page spec. Ask for the clauses that bind you past the first month and the ones that decide who owns the work. Then remember the unmarked-quotation note above before you paste any summary it wrote into something you publish.

The internal tool. The quote calculator that lives in a spreadsheet with macros, the booking-slot checker, the photo renamer. One user, cheap to break, nothing a customer touches. The moment a customer touches it, it moves to the other list.

The keep list

Anything that sends, posts or pays on its own. Anthropic's own small-business package is built so that "you approve before anything sends, posts, or pays". That line does not move because the model got smarter. Drafts go into a queue. Sends come from a person.

Any fact you would act on that came back at low effort. The default is high in Claude Code but medium in Cowork and on claude.ai, and at low the model answers from memory more. A quick question between meetings can return a confident answer that is a year old. Raise the effort or make it search.

Customer data, until your retention terms are written down. The zero-retention arrangement Anthropic announced, Enterprise Frontier Safeguards, arrives in phases from this autumn for enterprise customers. Until then, eligible customers can ask for zero data retention and everyone else is on their plan's standard terms. The safeguard classifiers also read everything the model reads, including files and connected tools. That is one page to write before the first customer record goes in, and no reason to wait on the rest.

Two harness changes worth making this week

If your own agent loop edits history between requests, it will start failing on Fable 5.1 for new accounts. Keep the conversation append-only and use turn-scoped system messages for per-turn reminders instead of injecting and deleting text:

messages.append({"role": "user", "content": tool_results}) messages.append({ "role": "system", "content": "First privately list what you need next; then request every item " "that doesn't depend on another's result in this one response.", "clear_at": "next_user_message", }) # beta header: mid-conversation-system-clear-at-2026-08-21

That nudge is also Anthropic's fix for the one-tool-call-per-turn behaviour. Earlier copies stay in the array; the API clears them once a later user message exists, so they cost no input tokens and do not invalidate the cache.

And if you want to see what it is doing during a long run, request the progress blocks. They come back empty under the default thinking.display of "omitted":

response = client.beta.messages.create( model="claude-fable-5-1", max_tokens=16000, thinking={"type": "adaptive", "display": "updates"}, betas=["thinking-display-updates-2026-08-18"], messages=messages, )

Any thinking block with non-empty text is then a status line you can show the user.

Plan mechanics, briefly

On Max plans and premium Team seats, Fable models are included up to half of your weekly usage limit and burn it faster than other models. On Pro and standard seats they run on usage credits from the first message. Anthropic says medium effort on 5.1 roughly matches Fable 5 at full effort, so the dial matters more than the model picker: default for drafts and reads, high for research and reviews, one long conversation per job so the cache does the work.

What the read caught in this post

The blog version of this piece, drafted the same way, compressed "four to five years" in the Millennium quote to "four years", and wrote a not-this-but-that sentence that our own style guide bans. The pipeline's review pass caught both before I read it. Both are behaviours from the second list, in the post about the second list. Both took a minute to fix. Neither would have been caught if the post had gone out on the strength of the capability list.

What did you hand Fable 5.1 first, and did it do anything the release notes warned about?